Rotation Invariance-Based HOG for Multi-posture Human Detection

The technology of human detection is evolving rapidly and it has been applied in various fields. Histograms of oriented gradients (HOG) are proven to be able to significantly outperform existing feature sets for human detection. This paper presents a foreground-based scalloped (FSR)-HOG features with rotation invariance for human detection. Firstly, extracting the FSRHOG human features on the foreground image, then training the positive and negative samples by SVM linear classifier. Experimental results show that the FSRHOG feature vector can retain the accuracy of HOG as well as improve the detection rate, and it is suitable for human detection with high motion complexity.